Add support for pickle protocol 4: It can be *much* faster

[jamadden]

Pickle protocol 4, introduced with Python 3.4, brings some nice enhancements. In particular, it turns out that the framing feature introduced in protocol 4 "with a potentially huge performance impact" really does have a big impact.

I'm using zodbpickle to gain access to pickle protocol 3 under Python 2 in RelStorage. Under Python 3, I use the builtin pickle and (under 3.4 and up) protocol 4.

In one test in RelStorage, protocol 4 takes about 5 seconds to handle an object, whereas protocol 3 takes about 30 seconds. (To workaround this in Python 2, I have to buffer a large object into memory; even using buffered IO from an SSD and with the data putatively in the buffer cache. Of course, this varies by system, but the slower the filesystem/stream the worse the problem).

It would be nice to have protocol 4 available to python 2.

I don't know how much work would be involved in the backport (and making sure that no_load stays supported). I personally don't have any concrete plans to do this work right now (maybe someday), but I thought I'd throw this feature request out there along with the motivating numbers in case it intrigues any one else.

[moreati]

I'm going to take a stab at this. My plan is to rebase on the CPython 3.7 code.

[moreati]

I've been unable to get the _pickle module from Python 3.7 working on Python 3.6. My efforts are in https://github.com/moreati/pikl/tree/zodbpickle-py37-backport. I'll try again with _pickle from 3.6 on 3.5.